Service level agreement.

Version 2.1 Last revised 2026-08-24 Part of the terms of service

1. The commitment

99.9% availability, per machine, per calendar month, on network reachability and power.

That is a budget of about 43 minutes in a thirty-day month. For comparison:

We commit to 99.9% and not more because a single dedicated machine cannot honestly promise more. Four nines needs redundancy across machines, and that is something you build on top of what we rent — not something we can sell you on one box.

2. What counts as downtime

A minute counts as downtime when your machine is unreachable from outside our network, or has lost power, for a reason within our control.

• Network: your machine's public address does not respond from outside our network, because of our routers, our switches, our uplinks or our transit.
• Power: the machine has lost power, or has been powered down by us.
• Hardware: a component we supplied has failed and the machine cannot run because of it. Counted from your report, and see clause 5.

Downtime is counted per machine. If you rent four servers and one is unreachable, that one has an incident; the other three do not.

3. How it is measured

• From outside the data centre, on the public address, not from a probe sitting in the same rack.
• At one-minute resolution. An incident is counted from the first failed check to the first successful one.
• Three consecutive failed checks are needed to open an incident, so a single dropped packet does not become an outage. Once open, all three minutes count.
• Measurement is continuous and automatic. It does not require you to have noticed.

The status page publishes what the probes recorded, component by component and site by site, and keeps every incident in its history — the same data this agreement is settled on, and the same log since September 2022.

4. What you get back

Time, not money. There is no account balance here to credit, so downtime is repaid in the only currency that is actually equivalent: more of the machine you rented.

5. Hardware replacement

Separate from the availability commitment, and worth stating on its own because it is the one you are most likely to use.

• Four-hour target from your report to functioning hardware, around the clock, on any machine.
• Spares — cards, PSUs, disks, and one complete chassis per model — are held in the same suite. A replacement is a walk down an aisle, not a courier booking.
• Your disks stay in place for a GPU, PSU or fan replacement. We touch storage only when storage is what failed, and we ask first.
• If we miss the four hours, the excess counts as downtime under clause 1 and is repaid under clause 4.

A failure caused by something you did — a kernel that will not boot, a firewall rule that locked you out, an fstab that panics the machine — is not a hardware failure. It is also not our problem to fix, because you have IPMI and can fix it yourself. See the documentation.

6. Scheduled work

Scheduled maintenance is excluded from the availability calculation, but only when it is genuinely scheduled:

• At least 72 hours' notice, on the contact handle you gave us.
• A stated window with a stated maximum duration.
• Between 01:00 and 05:00 in the local time of the data centre, unless a safety issue makes that impossible.

Work that overruns its stated window counts as downtime for the overrun. Emergency work — a failing power path, a security fix that cannot wait — is announced as early as we can and does count as downtime; we are not going to reclassify our own emergency as your scheduled maintenance.

We do not push firmware to a machine during a term. BIOS, BMC and VBIOS are set before the node enters the catalogue and applied again between tenants, never underneath you.

7. Exclusions

The following do not count as downtime. This is the complete list, and if something is not on it, it counts.

• Anything inside your operating system. A crashed kernel, a full disk, a misconfigured service, a firewall rule that locked you out, a driver you replaced. You have root; that side of the boundary is yours.
• Scheduled maintenance announced under clause 6, within its stated window.
• Suspension under the acceptable use policy.
• Non-payment. A term that has ended has no availability commitment.
• Denial-of-service attacks against you, beyond the mitigation capacity of the site. We filter rather than null-route your address, and we will keep filtering; we cannot commit to absorbing an attack of unlimited size.
• Failures of the public internet outside our network — a peer's outage, a submarine cable, a route leak by someone else.
• Force majeure: natural disaster, war, a government order, a utility failure affecting the whole facility beyond its generator runtime.
• Anything you asked us to do, including a reinstall or a power cycle you requested.

8. Disputing a measurement

If your own monitoring says a machine was down and ours does not, tell us. Bring timestamps with a time zone and, if you have them, the raw check results.

We will look at the same window from our side and tell you what we see. If our measurement was wrong, we correct it and apply clause 4 — being right is worth less to us than being trusted about it.

One thing worth knowing in advance: a machine unreachable from one place on the internet and reachable from everywhere else is usually a routing problem between your network and ours. It is still worth reporting — we can often fix it — but it will not appear as downtime in a measurement taken from outside.